Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

disable itrex format first (#998)

Weiwei395602b7b810
diff --git a/auto_round/utils/common.py b/auto_round/utils/common.py
index b1a5e18a6..6b1717e7b 100644
--- a/auto_round/utils/common.py
+++ b/auto_round/utils/common.py
@@ -88,8 +88,6 @@ class SupportedFormats:
"auto_round:gptqmodel",
"auto_round:auto_awq",
"auto_round:llm_compressor",
- "itrex",
- "itrex_xpu",
"fake",
"llm_compressor",
)
diff --git a/test/test_cpu/test_autoround.py b/test/test_cpu/test_autoround.py
index adadaa837..dd188e6ad 100644
--- a/test/test_cpu/test_autoround.py
+++ b/test/test_cpu/test_autoround.py
@@ -162,28 +162,6 @@ class TestAutoRound(unittest.TestCase):
print(result["results"]["lambada_openai"]["acc,none"])
self.assertGreater(result["results"]["lambada_openai"]["acc,none"], 0.35)
- def test_default(self):
- bits, group_size, sym = 4, 128, False
- autoround = AutoRound(
- self.model,
- self.tokenizer,
- bits=bits,
- group_size=group_size,
- sym=sym,
- iters=2,
- seqlen=2,
- dataset=self.llm_dataloader,
- )
- autoround.quantize()
-
- autoround.save_quantized(output_dir="./saved", inplace=False, format="itrex")
- try:
- import auto_gptq
- except:
- return
- if torch.cuda.is_available():
- autoround.save_quantized(output_dir="./saved", inplace=False)
-
def test_w4g1(self):
model_name = "/tf_dataset/auto_round/models/facebook/opt-125m"
bits, group_size, sym = 4, -1, True
diff --git a/test/test_cpu/test_autoround_acc.py b/test/test_cpu/test_autoround_acc.py
index 97211ade4..41b28e663 100644
--- a/test/test_cpu/test_autoround_acc.py
+++ b/test/test_cpu/test_autoround_acc.py
@@ -13,7 +13,6 @@ import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from auto_round import AutoRound # pylint: disable=E0401
-from auto_round.export.export_to_itrex.export import pack_model # pylint: disable=E0401
class LLMDataLoader: